Advancing Human Population Genomics with DNA Foundation Models

DNA foundation models offer a new approach to interpret genetic variation, but their potential in population-scale genomics remains untapped. We introduce a novel analytical framework that integrates a genomic foundation model with human population genomics studies. We employed the Evo2 DNA foundation model to systematically score the functional impact of a variant and haplotype across diverse cohorts including people with Alzheimers Disease Neuroimaging Initiative (ADNI), the Human Pangenome Project, and the UK Biobank. As proof-of-concept, the analysis of the APOE locus confirmed the approachs validity, with model-derived scores can help to prioritize putatively functional variants and quantify effects of both variants and haplotype onto Alzheimers Disease susceptibility or associated endophenotypes, including cognitive performance, brain structural change and amyloid load. Specifically, scoring multi-ancestry assembly sequences from the Human Pangenome Project revealed, for the first time, that genetic variation could nicely explains ancestry-specific differences in APOE expression and the impact of APOE-{varepsilon}4 on Alzheimers disease risk. Overall, this study provides a scalable framework for mapping functional genetic variation, complementing conventional population-genomics approaches, enabling better interpretation of genetic effects in complex genomic regions at population scale.